from abc import ABC, abstractmethod from contextlib import contextmanager, nullcontext from copy import copy from dataclasses import dataclass from functools import partial, wraps from typing import Any, Callable, cast, Dict, List, Optional, Set, Tuple, Union from functorch import make_fx import torch import torch.distributed as dist # We need to import _functional_collectives to trigger op registration import torch.distributed._functional_collectives import torch.nn as nn import torch.utils._pytree as pytree from torch import fx from torch._decomp.decompositions import native_layer_norm_backward from torch._subclasses.fake_tensor import FakeTensorMode from torch.distributed._spmd.data_parallel import gradients_tagging from torch.distributed._spmd.parallel_mode import ( DataParallel, DTensorExpandMode, ParallelMode, ) from torch.distributed._tensor import Placement from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo, CodeGen from torch.nn.utils import stateless from torch.nn.utils._named_member_accessor import NamedMemberAccessor class Override(ABC): r""" Override the tracing and transformation behavior of :meth:`~torch.distributed._spmd.compile`. This is useful when any part of the model is not traceable or if you prefer to not trace it due to any reason. More specifically, users can implement :meth:`torch.distributed._spmd.Override.replacement` to replace an original submodule with the return new submodule. The new submodule contains operations that users preferred to be traced, which simply be a dummy placeholder operator. After tracing, users can implement :meth:`torch.distributed._spmd.Override.transform` to transform the traced graph, where the dummy placeholder operator serves as an anchor to insert new sub-graphs. """ @abstractmethod def replacement(self, fqn: str, orig_submodule: torch.nn.Module) -> torch.nn.Module: r""" Implement this method to return a new :class:`nn.Module` instance to replace the ``orig_submodule`` argument in the model. This helps if ``orig_submodule`` is not traceable or should not be traced. Args: fqn (str): fully quantified name of the submodule. orig_submodule (class:`nn.Module`): original submodule instance to replace. Returns: A new :class:`nn.Module` instance to replace the original one. """ pass @abstractmethod def transform( self, gm: fx.GraphModule, flat_state: List[torch.Tensor], ) -> fx.GraphModule: r""" Given a DTensor-expanded graph and sharding schema for every node, conduct additional transformation for the sub-graph from the :class:`nn.Module` returned by :meth:`torch.distributed._spmd.Override.replacement` if necessary. Args: gm (:class:`fx.Graph`): a DTensor-expanded graph. flat_state (List[str, :class:`Tensor`]): a reference to the list of flattened state. The elements in ``flat_state`` map to the first ``len(flat_state)`` placeholders in the graph. The transformation can add state to or remove state from ``flat_state`` as long as it keeps ``flat_state`` and the placeholders consistent. Returns: The :class:`fx.Graph` after transformation. """ pass class _PyTreeCodeGenOutputsOnly(_PyTreeCodeGen): # pyre-ignore[3] def process_inputs(self, *args: Any) -> Any: return args # pyre-ignore[2, 3] def gen_fn_def(self, free_vars, maybe_return_annotation): return CodeGen.gen_fn_def(self, free_vars, maybe_return_annotation) def _to_caller_flattened_graph_module(gm: torch.fx.GraphModule) -> torch.fx.GraphModule: """Move the responsibility of flattening the input arguments from the graph module to the caller. Example: output = gm(my_struct) gm = gm(to_caller_flattened_graph_module) output = gm(*pytree.flatten(my_struct)[0]) """ # pyre-ignore[16] gm._graph._codegen = _PyTreeCodeGenOutputsOnly( pytree_info=_PyTreeInfo( # pyre-ignore[6] orig_args=None, # type: ignore[arg-type] # pyre-ignore[6] in_spec=None, # type: ignore[arg-type] # pyre-ignore[16] out_spec=gm._graph._codegen.pytree_info.out_spec, ) ) gm.recompile() return gm # Use a dtensor expand mode for now to preserve the old behavior # and avoid breaking existing code dtensor_expand_mode = DTensorExpandMode() def _override_placements(t: torch.Tensor, placements: List[Placement]): global dtensor_expand_mode dtensor_expand_mode._placements_override[id(t)] = placements @contextmanager def _rematerialize_optimizer( opt: torch.optim.Optimizer, named_states: Dict[str, Any], params: Dict[str, nn.Parameter], ): assert opt is not None # update opt.state with proxy tensors orig_states = copy(opt.state) for n in named_states: # opt.state's key type is string, but optimizer uses Parameter as keys opt.state[params[n]] = named_states[n] # type: ignore[index] # FIXME: support multiple parameter groups param_group = opt.param_groups[0] orig_params = param_group["params"] param_group["params"] = params.values() try: yield finally: param_group["params"] = orig_params opt.state = orig_states aten = torch.ops.aten # pyre-ignore @contextmanager def _enable_compile(): # The return value of torch._utils.is_compiling changes optimizer behavior. # We need that function to return True to include optimizer in the graph. # See: https://github.com/pytorch/pytorch/blob/a524123c91ab399c9dd6882c1189596dd77e7734/torch/optim/optimizer.py#L41 def f_true(): return True orig_is_compiling_code = torch._utils.is_compiling.__code__ torch._utils.is_compiling.__code__ = f_true.__code__ try: yield finally: torch._utils.is_compiling.__code__ = orig_is_compiling_code def _foreach_add_decomp(self, other, alpha=1): self_updated = aten._foreach_add.List(self, other, alpha=alpha) for s, s_u in zip(self, self_updated): s.copy_(s_u) def _foreach_unaop_decomp(op, self): self_updated = op(self) for s, s_u in zip(self, self_updated): s.copy_(s_u) def _foreach_binop_list_decomp(op, self, other): self_updated = op(self, other) for s, s_u in zip(self, self_updated): s.copy_(s_u) def _foreach_binop_scalar_decomp(op, self, scalar=1): self_updated = op(self, scalar) for s, s_u in zip(self, self_updated): s.copy_(s_u) def _foreach_addcop_scalar_decomp(op, self, tensor1, tensor2, scalar=1): self_updated = op(self, tensor1, tensor2, scalar) for s, s_u in zip(self, self_updated): s.copy_(s_u) def _fused_adam_decomp( self, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps, *, lr=1, beta1=1, beta2=1, weight_decay=1, eps=1, amsgrad=True, maximize=True, grad_scale=None, found_inf=None, ): orig_tuple = (self, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs) updated_tuple = aten._fused_adam.default( self, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps, lr=lr, beta1=beta1, beta2=beta2, weight_decay=weight_decay, eps=eps, amsgrad=amsgrad, maximize=maximize, grad_scale=grad_scale, found_inf=found_inf, ) for idx, (orig, updated) in enumerate(zip(orig_tuple, updated_tuple)): if idx == 1: # skip gradient copying as we don't need to copy gradients back continue for o, u in zip(orig, updated): o.copy_(u) SPMD_DECOMP_TABLE = { aten._foreach_add_.List: _foreach_add_decomp, aten._foreach_add_.Scalar: partial( _foreach_binop_scalar_decomp, aten._foreach_add.Scalar ), aten._foreach_addcdiv_.Scalar: partial( _foreach_addcop_scalar_decomp, aten._foreach_addcdiv.Scalar ), aten._foreach_addcmul_.Scalar: partial( _foreach_addcop_scalar_decomp, aten._foreach_addcmul.Scalar ), aten._foreach_div_.List: partial( _foreach_binop_list_decomp, aten._foreach_div.List ), aten._foreach_mul_.Scalar: partial( _foreach_binop_scalar_decomp, aten._foreach_mul.Scalar ), aten._foreach_div_.Scalar: partial( _foreach_binop_scalar_decomp, aten._foreach_div.Scalar ), aten._foreach_neg_.default: partial( _foreach_unaop_decomp, aten._foreach_neg.default ), aten._foreach_reciprocal_.default: partial( _foreach_unaop_decomp, aten._foreach_reciprocal.default ), aten._foreach_sqrt_.default: partial( _foreach_unaop_decomp, aten._foreach_sqrt.default ), aten._foreach_sub_.Scalar: partial( _foreach_binop_scalar_decomp, aten._foreach_sub.Scalar ), aten._fused_adam_.default: _fused_adam_decomp, aten.native_layer_norm_backward.default: native_layer_norm_backward, } DEDUP_TARGETS: Set[torch._ops.OpOverload] = { torch.ops.c10d_functional.all_reduce.default, torch.ops.c10d_functional.wait_tensor.default, } def _dedup_collectives(gm: fx.GraphModule) -> fx.GraphModule: args_to_node: Dict[Tuple[Any, ...], fx.Node] = {} for node in gm.graph.nodes: # replace all args with the results from the first unique comm op args, _ = pytree.tree_flatten(node.args) if node.target in DEDUP_TARGETS: args_key = (node.target, *args) unique_node = args_to_node.get(args_key, None) if unique_node is None: # first time seeing this combination, remember it args_to_node[args_key] = node else: # the current node is a duplicate, replace it node.replace_all_uses_with(unique_node) gm.graph.erase_node(node) gm.recompile() return gm @dataclass class _CompiledResult: gm: fx.GraphModule mod: nn.Module opt: Optional[torch.optim.Optimizer] flat_state: List[torch.Tensor] def _compile( func: Callable, module_override: Optional[List[Override]], parallel_mode: ParallelMode, *args: Any, **kwargs: Any, ) -> _CompiledResult: # 1. Extract nn.Module and Optimizer from args and kwargs # FIXME(@mrshenli): support multiple nn.Module instances # FIXME(@mrshenli): support multiple Optiimzer instances # FIXME(@mrshenli): need to broadcast model to sync parameters mod, opt = None, None for arg in pytree.tree_flatten(list(args) + list(kwargs.values()))[0]: if isinstance(arg, nn.Module): assert mod is None, "Only support single nn.Module for now" mod = arg if isinstance(arg, torch.optim.Optimizer): assert opt is None, "Only support single Optimizer for now" opt = arg assert mod is not None, "Couldn't find nn.Module instances from the arguments." # 2. Override target submodules (e.g., MoE) with dummy replacements if module_override: accessor = NamedMemberAccessor(mod) def swap(fqn_prefix: str, module: torch.nn.Module) -> None: for override in module_override: # type: ignore[union-attr] for name, child in module.named_children(): if len(name) == 0: continue fqn = fqn_prefix + "." + name if fqn_prefix != "" else name new_child = override.replacement(fqn, child) if id(new_child) == id(child): swap(fqn, new_child) else: accessor.swap_submodule(fqn, new_child) swap("", mod) # 3. Trace statelss version of the train_step params = dict(mod.named_parameters(remove_duplicate=False)) buffers = dict(mod.named_buffers(remove_duplicate=False)) named_states = {} if opt is not None: # Pass named_states instead of opt.state to stateless_func, because # the later uses nn.Parameter as key. During tracing, we need to # make sure optimizers can find the states using proxy tensors. for n, p in params.items(): if p in opt.state: # opt.state's key type is string, but optimizer uses # Parameter as keys named_states[n] = opt.state[p] # type: ignore[index] is_data_parallel_mode = isinstance(parallel_mode, DataParallel) # Lift states and parameters as function arguments so that make_fx # can trace operations applied to them. def stateless_func(func, params, buffers, named_states, args, kwargs): with stateless._reparametrize_module( mod, {**params, **buffers} ), _rematerialize_optimizer( opt, named_states, params ) if opt else nullcontext(): # For DataParallel mode, install hooks first to tag the gradients with gradients_tagging(params) if is_data_parallel_mode else nullcontext(): ret = func(*args, **kwargs) # make sure updated parameters are returned return ret, list(mod.parameters()), list(named_states.values()) # type: ignore[union-attr] # FIXME: Using symbolic tracing to work around in DTensor expand mode. # Otherwise it hits shape mismatch error, as we use local inputs to # trace local graph and use DTensor to expand operators, where # DTensor's shape is the global shape. tracing_mode = "fake" if is_data_parallel_mode else "symbolic" if is_data_parallel_mode: fake_mode = FakeTensorMode() data_parallel_mode = cast(DataParallel, parallel_mode) def _get_full_batch_arg(arg: torch.Tensor) -> torch.Tensor: # since compilation happens in the first iteration and we # receives mini-batch input, convert them to full batch # fake tensor input first for data parallel sharding # propagations fake_arg = fake_mode.from_tensor(arg) arg_dims = [1] * arg.ndim # expand the tensor to full batch size on its batch dim arg_dims[data_parallel_mode.input_batch_dim] *= dist.get_world_size() return fake_arg.repeat(arg_dims) args = pytree.tree_map_only( torch.Tensor, _get_full_batch_arg, args, ) kwargs = pytree.tree_map_only( torch.Tensor, _get_full_batch_arg, kwargs, ) with _enable_compile(), torch.autograd.detect_anomaly(check_nan=False): # FIXME(@mrshenli): functionalization does not work for our use # case yet. Use explicit decompositions for foreach ops. # Remove this when the following issue is addressed. # Issue: https://github.com/pytorch/pytorch/issues/97852 gm = make_fx( partial(stateless_func, func), tracing_mode=tracing_mode, decomposition_table=SPMD_DECOMP_TABLE, _allow_non_fake_inputs=False, )(params, buffers, named_states, args, kwargs) params_and_buffers: Dict[str, Union[torch.Tensor, nn.Parameter]] = { **params, **buffers, } # 4. parallel mode to expand a single device graph to a distributed graph gm = parallel_mode.partition( gm, mod, opt, params_and_buffers, named_states, args, kwargs, ) # 5. Move the responsibility of flattening the input arguments from the # graph module to the caller. This serves two purposes: # - Transformations that add/remove state need to manipulate a state # container that maintains the state tensors in the same order as they # appear in graph placeholders. # - Reduced runtime cost. The state container is only flattened once upfront. flat_state, _ = pytree.tree_flatten([params_and_buffers, named_states]) gm = _to_caller_flattened_graph_module(gm) # 6. dedup comm operators. # The duplication could come from DTensor args and kwargs redistribution. # Suppose one operator produces a Partial gradient tensor and model # parameters are replicated. In this case, every optimizer operation using # that Partial gradient tensor would trigger an allreduce. This is becuase # DTensor only has local information on individual tensor/operator, which is # not sufficient to detect duplications in the graph. This situation can # also happen when inserting FSDP allgather if a parameter is used multiple # times in the forward method. # TODO(@mrshenli): @yifuwang has a suggestion of conducting expansion and # dedup at tracer-level to avoid multiple graph passes. gm = _dedup_collectives(gm) # 7. Replace previously inserted dummy ones with real graphs. if module_override: for override in module_override: gm = override.transform(gm, flat_state) return _CompiledResult(gm, mod, opt, flat_state) # Note that the Python convention of __dict__ requires the key to be str. # TODO: ensure the key is unique. COMPILED_OBJECT_KEY = "_compiled_obj" def compile( module_override: Optional[List[Override]] = None, gm_transformation: Optional[Callable[[fx.GraphModule], fx.GraphModule]] = None, parallel_mode: Optional[ParallelMode] = None, ): r""" Compile and optimize a callable, which can be a train step within a training loop. This method will extract :class:`nn.Module` and :class:`torch.optim.Optimizer` instances from the input arguments and trace operations applied to their parameters and states. Args: module_override (Optional[List[Override]]): a list of Override instances that will be applied to the module in order. The :class:`Override` objects provide :class:`nn.Module` replacements during tracing and a graph transformation function after tracing. (Default: ``None``) gm_transformation (Optional[Callable[fx.GraphModule, fx.GraphModule]]): a callback that will be called after the original callable is compiled and distributed (usually after the first iteration) to transform the compiled GraphModule into a new optimized one. parallel_mode (Optional[ParallelMode]): a :class:`ParallelMode` object that specifies how to parallelize the callable. Each ParallelMode would have its own strategy to partition the model and the captured graph (Default: ``None``) """ def inner(func: Callable): @wraps(func) def wrapper(*args, **kwargs): last_train_step = kwargs.pop("last_train_step", False) if kwargs else False first_iter = False # Put the COMPILED_OBJECT_KEY in ``wrapper`` instead of ``func`` as # ``wrapper`` is the one that users will get. compiled_obj = wrapper.__dict__.get(COMPILED_OBJECT_KEY, None) if compiled_obj is None: first_iter = True global dtensor_expand_mode mode: ParallelMode = ( dtensor_expand_mode if parallel_mode is None else parallel_mode ) compiled_obj = _compile(func, module_override, mode, *args, **kwargs) wrapper.__dict__[COMPILED_OBJECT_KEY] = compiled_obj flat_inps = compiled_obj.flat_state + pytree.tree_flatten([args, kwargs])[0] with torch.no_grad(): # N.B.: we don't need autograd as backward has already been # captured in the graph. if first_iter and gm_transformation: # TODO: SPMD should provid a default and configurable # transformation. compiled_obj.gm = gm_transformation(compiled_obj.gm) if not last_train_step: output = compiled_obj.gm(*flat_inps)[0] else: # This is the last train step. Call IterGraphModule.forward() # with the `last_iter` argument and catch the exception in # case the compiled_obj is not wrapped with IterGraphModule. try: output = compiled_obj.gm(*flat_inps, last_iter=last_train_step)[ 0 ] except TypeError as e: if "last_iter" not in str(e): raise e output = compiled_obj.gm(*flat_inps)[0] return output return wrapper return inner